AI beats human at a game of physical skill for the first time

An artificially intelligent robot has beaten humans at a Labyrinth marble game, the first time a machine has shown it can master both physical dexterity and learning.

The game, which involves manipulating a marble around a series of obstacles while stopping the ball tumbling into multiple holes, requires skill, patience and a delicate touch.

Currently, Sweden’s Lars Goran Danielsson, who has played the game for 35 years, holds the record for completing the maze in 15.41 seconds.

But a robot called CyberRunner, built by Swiss researchers at ETH Zurich, clocked a time of 14.48 seconds after just six hours of practice.

Raffaello D’Andream Professor of Dynamic Systems and Control at ETH Zurich said: “It’s not just about beating humans at a game. It’s beating humans at a game that requires direct physical skill and dexterity.

“The first model that allowed a machine to beat a human at a game was instrumental. Now what we are doing here is not just to change the type of challenge. We are adding dimensions, adding multiple boundaries to be broken.”

In the Labyrinth game, which was first brought to market by British company Brio in 1946, he the movement of the marble is controlled through two handles which tip or tile the board.

While a relatively straightforward game, it requires fine motor skills and spatial reasoning abilities, and humans require a great amount of practice to become proficient at the game.

The robot uses two motors to move the handles and watches its progress through a camera mounted above the board.

Just like humans, CyberRunner learns through experience, recognising which strategies and behaviours are more promising, so that it gets better run after run, using model-based reinforcement learning.

Computer scientists were intrigued to find that CyberRunner’s first instinct was to cheat, finding quicker ways to navigate the labyrinth by skipping parts of the maze.

Thomas Bi, a doctoral candidate at ETH Zurich, said: “Interestingly, during the learning process, CyberRunner naturally discovered shortcuts. It found ways to ’cheat’ by skipping certain parts of the maze.

“We had to step in and explicitly instruct it not to take any of those shortcuts.”

AI has a long history of beating humans, winning its first chess tournament in 1976, and besting grandmaster Garry Kasparov in 1997.

In 2016, Google DeepMind’s AlphaGo AI system defeated world Go champion Lee Sedol, inventing new moves that had not been seen before.

Experts said the new breakthrough marked the world’s first physical application of AI in which it beat a human.

Bi added: “The AI robot outperforms the previously fastest recorded time, achieved by an extremely skilled human player, by over six per cent.”

The team plans to release the code and hardware details for the CyberRunner shortly.

A paper describing the work was posted on the arxiv.org preprint server.